Adaptive fuzzy control for a class of unknown fractional-order neural networks subject to input nonlinearities and dead-zones

Adaptive fuzzy control for a class of unknown fractional-order neural networks subject to input nonlinearities and dead-zones

Accepted Manuscript Adaptive Fuzzy Control for a Class of Unknown Fractional-Order Neural Networks Subject to Input Nonlinearities and Dead-Zones Hen...

1MB Sizes 1 Downloads 48 Views

Accepted Manuscript

Adaptive Fuzzy Control for a Class of Unknown Fractional-Order Neural Networks Subject to Input Nonlinearities and Dead-Zones Heng Liu, Shenggang Li, Hongxing Wang, Yeguo Sun PII: DOI: Reference:

S0020-0255(18)30326-8 10.1016/j.ins.2018.04.069 INS 13612

To appear in:

Information Sciences

Received date: Revised date: Accepted date:

6 March 2017 13 April 2018 21 April 2018

Please cite this article as: Heng Liu, Shenggang Li, Hongxing Wang, Yeguo Sun, Adaptive Fuzzy Control for a Class of Unknown Fractional-Order Neural Networks Subject to Input Nonlinearities and Dead-Zones, Information Sciences (2018), doi: 10.1016/j.ins.2018.04.069

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Adaptive Fuzzy Control for a Class of Unknown Fractional-Order Neural Networks Subject to Input Nonlinearities and Dead-Zones

CR IP T

Heng Liua,b , Shenggang Lia,∗, Hongxing Wangb , Yeguo Sunb a College

of Mathematics and Information Science, Shaanxi Normal Universtiy, Xi’an 710119, China b Department of Mathematics and Computational Science, Huainan Normal University, Huainan 232038, China

AN US

Abstract

This paper presents an adaptive fuzzy control (AFC) for uncertain fractionalorder neural networks (FONNs) with input nonlinearities and unmodeled dynamics. System uncertainties and unknown parts of the nonlinear input are approximated by fuzzy logic systems (FLSs). Based on some proposed stability analysis criteria for fractional-order systems (FOSs), an AFC is designed to guar-

M

antee the asymptotic stability of the controlled system. Fractional-order adaptive laws (FOALs) are constructed to update adjustable parameters of FLSs.

ED

Our method can be used to control FONNs with/without sector nonlinearities in control inputs. It also allows us to generalize many existing control methods that are valid for integer-order neural networks to FONNs by using the proposed

PT

method. Finally, the effectiveness of the proposed method is demonstrated by simulation results.

Keywords: Adaptive fuzzy control, sector nonlinearity, fractional-order neural

CE

network, dead-zone

AC

1. Introduction Up to now, neural networks (NNs) have received increasing attention in

many fields [2, 13, 17, 28, 29, 34, 35, 45]. The NN that will be considered in this paper is called fractional-order neural network (FONN), which is an extension of ∗ Corresponding author, College of Mathematics and Information Science, Shaanxi Normal Universtiy, Xi’an 710119, China; Email: [email protected]

Preprint submitted to Elsevier

April 23, 2018

ACCEPTED MANUSCRIPT

ordinary NNs and have also been well studied [10, 28, 53] because the fractionalorder derivative provides a powerful technique for describing hereditary and memory properties of the system. FONNs have two advantages: one is that the fractional calculus has infinite memory, the other is that it has one more on the control and synchronization of FONNs.

CR IP T

freedom [3, 27]. Please see [10, 14] and some references therein for recent results Fuzzy logic systems (FLSs) with linguistic information have been effectively

utilized in the control of unknown nonlinear systems [4, 6, 7, 9, 11, 18, 20, 21, 24,

25, 26, 30, 33, 36, 42, 47, 48, 49, 50]. However, there are few works that studied the adaptive fuzzy control (AFC) of FONNs because it is a challenging work to

AN US

check the stability by using a fractional-order stability criterion. On the other

hand, we notice that inputs (particularly, nonlinear inputs), outputs, and state constraints, which are reflected by physical stoppages, dead-zones, backlash, saturation, and safety specifications, are very common in most physical systems, and nonlinear inputs usually lead to instability of the controlled system [7, 37, 46]. Thus control of nonlinear systems with nonlinear control inputs has received increasing attention [7, 22, 41, 46, 51]. However, the stability of FONNs cannot

M

be established by using polynomial criteria [16, 29] because it is impossible to check the stability of FONNs simply by looking for its dominant roots. In [39],

ED

an adaptive switching control method was proposed for fractional-order systems (FOSs) subject to input nonlinearities, but the method used in the stability analysis is the integer-order Lyapunov method. Therefore, a fractional-order analysis method for stability of FONNs with nonlinear inputs is expected.

PT

The present paper will design an AFC for FONNs subject to input nonlinearities and dead-zones. There are three reasons that motivate us to study this

CE

control problem. Firstly, since system models are partly known for most works about the control and synchronization of FONNs, it is advisable to develop control methods for FONNs with fully unknown system structures. Secondly, few

AC

results have been given for discussion of the control of FOSs with input nonlinearities. Thirdly, the fractional-derivative of the quadratic Lyapunov function V (t) = 12 eT (t)P e(t) (which is usually be utilized to discuss the system’s stability) is given by the following complex form: C α 0 Dt V

(t) =

∞ X

k=0

Γ(1 + α) α−k C k D e(t)C e(t). 0 Dt Γ(1 + k)Γ(1 − k + α) 0 t 2

ACCEPTED MANUSCRIPT

Clearly, it is difficult to verify the stability of the FOSs by using this Lyapunov function [40, 43]. Therefore, new ideas or methods are needed. Compared with existing results (for examples, [1, 19, 39]), our study has the following contributions:

CR IP T

1) An AFC is proposed for uncertain FONNs with fully unknown system

structures, where the new results (see Lemmas 4 and 5) can be expediently used to check the stability of FONNs by using fractional Lyapunov methods.

2) Input nonlinearities with unknown gain reduction tolerances are considered, and the values of gain reduction tolerances and the upper bound of unknown nonlinear functions are estimated by FLSs.

AN US

3) A FOAL is designed to estimate adjustable parameter of the FLS online. The stability of the closed-loop system is proved strictly by using fractional Lyapunov stability criterion, where the asymptotic convergence of system variables can be guaranteed. Moreover, some controllers which are valid for NNs can be extended to control FONNs based on our results.

M

2. Preliminaries and problem statement 2.1. Fractional calculus

The α-th fractional integral and α-th fractional derivative are defined by the

ED

following equalities (1) and (2)

PT

−α 0 Dt f (t) =

C α 0 Dt f (t)

=

1 Γ(α)

Z

t

(t − ξ)α−1 f (ξ)dξ,

0

1 Γ(1 − α)

Z

0

t

(t − ξ)−α f 0 (ξ)dξ,

(1)

(2)

AC

CE

respectively, where Γ is the ordinary Gamma function. It follows from (2) that L





C α 0 Dt f (t)

= sα F (s) −

n−1 X

sα−k−1 f (k) (0)

(3)

k=0

where F (s) = L {f (t)}.

Definition 1. [38]. The mapping Eα,γ : C −→ C, defined by Eα,γ (ζ) =

∞ X

k=0

3

ζk , Γ(αk + γ)

(4)

ACCEPTED MANUSCRIPT

is called Mittag-Leffler function, where C is the set of all complex numbers, and α, γ > 0. Taking Laplace transform on (4) gives [38] sα−γ . sα + a

(5)

CR IP T

L {tγ−1 Eα,γ (−atα )} =

Definition 2. [19]. A strictly increasing function h : [0, t) −→ [0, ∞) satisfying h(0) = 0 is called a class-K function.

AN US

Definition 3. [19] A continuous function z(t) ∈ Rn (i.e. z : [0, ∞) −→ Rn ) is said to be Mittag-Leffler stable if the Euclidean norm of z(t) satisfies b

kz(t)k ≤ {g[z(0)]Eα (−κtα )} ,

(6)

where κ > 0, b > 0, g : Rn −→ [0, ∞) is locally Lipschitz on {z(t) (i.e. on {z(t) | t ∈ [0, ∞)}) and also satisfies g(00) = 0, and 0 = [0, · · · , 0]T ∈ Rn .

M

Lemma 1. [19] Suppose that

C α 0 Dt h(t)

= f (t, h(t))

(7)

ED

with 0 being an equilibrium point. If there exist some class-K functions g1 , g2 , g3 and a Lyapunov function V (t, h(t)) such that

PT

g1 (kh(t)k) ≤ V (t, h(t)) ≤ g2 (kh(t)k), C α 0 Dt V

(t, h(t)) ≤ −g3 (kh(t)k),

(8) (9)

then lim h(t) = 0. t→∞

AC

CE

Lemma 2. [27] If z(t) ∈ Rn is a smooth function, then 1C α T α D z (t)Bz(t) ≤ z T (t)B C 0 Dt z(t), 20 t

where B ∈ Rn×n is positive definite.

Lemma 3. If

C α 0 Dt z(t)

≤ 0, then z(t) ≤ z(0) (t ∈ [0, +∞)).

4

(10)

ACCEPTED MANUSCRIPT

α Proof. Let h(t) = −C 0 Dt z(t) (t ∈ [0, +∞)). Then, for each t ∈ [0, +∞),

h(t) ≥ 0 and

C α 0 Dt z(t)

+ h(t) = 0.

(11)

Z(s) =

z(0) H(s) − α . s s

From (12) one has −α z(t) = z(0) − C 0 Dt h(t),

which implies z(t) ≤ x(0) (t ∈ [0, +∞)).

AN US

Lemma 4. Let V1 (t) = 12 x2 (t) + 12 y 2 (t). If

CR IP T

It follows from (11) that

C α 0 Dt V1 (t)

where k > 0. Then

≤ −kx2 (t),

x2 (t) ≤ 2V1 (0)Eα (−2ktα ). Proof. Using the operator

C −α 0 Dt

M

ED

−α 2 x2 (t) ≤ 2V1 (0) − 2k C 0 Dt x (t),

and thus

(13)

(14) (15)

to both sides of (14) gives

−α 2 V1 (t) − V1 (0) ≤ −k C 0 Dt x (t).

By (16),

(12)

−α 2 x2 (t) + m(t) = 2V1 (0) − 2k C 0 Dt x (t),

(16)

(17)

(18)

PT

where m(t) ≥ 0. It follows from (18) that X2 (s) = 2V1 (0)

sα−1 sα −2 α M (s), + 2k s + 2k



(19)

CE

where X2 (s) = L {x2 (t)}. According to (5), (19) can be solved as x2 (t) = 2V1 (0)Eα (−2ktα ) − 2m(t) ∗ [t−1 Eα,0 (−2ktα )],

(20)

AC

where ∗ is the convolution operator. Thus Eα,0 (−2ktα ) ≥ 0. By (20), we know that (15) holds.

Lemma 5. Let V2 (t) = 21 xT (t)Q1 x(t)+ 12 z T (t)Q2 z(t), where x(t) and z(t) ∈ Rn are smooth functions, and Q1 , Q2 ∈ Rn×n are two positive definite matrices. If C α 0 Dt V2 (t)

≤ −h0 xT (t)Q3 x(t),

where Q3 > 0, h0 > 0, then lim kx(t)k = 0. t→∞

5

(21)

ACCEPTED MANUSCRIPT

2.2. Description of FLS

j∈J

(22)

CR IP T

Usually, a FLS is modeled by [7, 8, 20, 21, 50]: P θj (t)µj (x(t)) j∈J P , fˆ(x(t)) = µj (x(t))

where fˆ (a Lipschitz continuous mapping from a compact subset Ω ⊆ Rn to n Q the real line R) is called the output, x is called the input, J = Fi , Fi i=1

consists of Ni fuzzy numbers (1 ≤ i ≤ n), µj (a mapping from Rn to the

closed unit interval [0, 1] ⊆ R which is defined relying on the above fuzzy numbers) is called the membership function of rule j

(j ∈ J), and θj is

We may identify

AN US

called the centroid of the j-th consequent set (j ∈ J).

J with {1, 2, · · · , N } for convenience. Denote θ(t) = [θ1 (t), · · · , θN (t)]T and ψ(x(t)) = [q1 (x(t)), q2 (x(t)), · · · , qN (x(t))]T , where qj (called the j-th fuzzy ba-

sis function, j ∈ J) is a continuous mapping (and thus, ψ : Ω −→ RN is continuous) defined by

θ (t) qj (x(t)) = P j . µs (x(t))

M

s∈J

Consequently, system (22) can be rewritten as

(23)

ED

fˆ(x(t)) = θT (t)ψ(x(t)).

PT

2.3. Problem statements

Consider the following FONNs: = −ci xi (t)+

CE

C α 0 Dt xi (t)

n X

aij fj (xj (t))+Ii +

j=1

n X j=1

gij ϕj (uj (t)), i = 1, · · · , n, (24)

in which n is the number of units, xi (t) represents the i-th unit’s state, aij

AC

corresponds to the constant connection weight, ci is the rate, gij is the unknown constant control gain, ϕj (uj (t)) represents a nonlinear input function, and Ii corresponds to the external input. Suppose that fi is Lipschitz-continuous.

Thus, system (24) has an unique solution [15]. Let x(t) = [x1 (t), · · · , xn (t)]T , ϕ(u(t)) = [ϕ1 (u1 (t)), · · · , ϕn (un (t))]T , C =

diag(c1 , · · · , cn ), I = [I1 , · · · , In ]T ,

6

ACCEPTED MANUSCRIPT

g11 · · · g1n  . .. .. . G= . .  . gn1 · · · gnn Then FONN (24) can be C α 0 Dt x(t)





a11  .  , A =  .  .  an1 rewritten as

··· .. . ···

 a1n ..   . . ann

CR IP T



= −Cx(t) + Gϕ(u(t)) + Af (x(t)) + I.

(25)

Assumption 1. The unknown control gain matrix G in system (25) is positive definite. 2.4. Description of the input nonlinearities

The input nonlinearities usually exist due to the limitations of actuators.

AN US

The model of ϕi (ui (t)) is defined as in [7, 39]:   hi+ (ui (t))(ui (t) − µ∗+ ), ui (t) > µ∗+ ,    ϕi (ui (t)) = 0, − µ∗− ≤ ui (t) ≤ µ∗+ ,     h (u (t))(u (t) + µ∗ ), u (t) < −µ∗ , i− i i i − −

(26)

where hi+ (ui (t)) > 0 and hi− (ui (t)) > 0 are nonlinear functions, and µ∗+ and

M

µ∗− are known constants. Based on the results in [7] and [39], it is reasonable to assume that ϕi (ui (t)) satisfies the followings:

ED

(ui (t) − µ∗+ )ϕi (ui (t)) ≥ d∗i+ (ui (t) − µ∗+ )2 ,

(ui (t) + µ∗− )ϕi (ui (t)) ≥ d∗i− (ui (t) + µ∗− )2 ,

(27)

where d∗i+ and d∗i− are “gain reduction tolerances”. Denote di = min{d∗i+ , d∗i− }.

PT

A dead-zone input nonlinearity is presented in Figure 1.

CE

Assumption 2. The input nonlinearity ϕi (ui (t)) is unknown, and the gain reduction tolerances, d∗i+ and d∗i− , are also unknown (furthermore, di is unknown).

AC

Remark 1. In fact, the model (26) was also used in [12, 31, 32, 39]. However, in above literature, the exact values of gain reduction tolerances should be known in advance. 3. Design of AFC and stability analysis Denote G−1 = Λ. As G > 0, Λ > 0. Multiplying both sides of (25) by Λ

gives α ΛC 0 Dt x(t) = −ΛCx(t) + ϕ(u(t)) + ΛAf (x(t)) + ΛI.

7

(28)

ACCEPTED MANUSCRIPT

Note that f (x(t)) and Λ are unknown, we can rewrite (28) into the following from: α ΛC 0 Dt x(t) = −ΛCx(t) + $(t) + ϕ(u(t)),

(29)

CR IP T

where $(t) = ΛAf (x(t)) + ΛI is an unknown nonlinear function. It is easy to know that $(x(t)) = ΛAf (x(t)) + ΛI is a Lipschitz continuous function, thus we can give the following assumption.

Assumption 3. There exists a continuous function $ ¯ i (x(t)) such that |$i (x(t))| ≤ d$ ¯ i (x(t)), where d = min {di }.

AN US

1≤i≤n

(30)

Remark 2. Assumption 3 is not restrictive because the upper bound function $ ¯ i (x(t)) is assumed to be unknown. At the same time, $ ¯ i (x(t)) always exists since $(x(t)) is continuous. We can employ FLS (23) to approximate $ ¯ i (t) by

M

ˆ $ ¯ i (t) = θiT (t)ψi (t),

(31)

ED

where ψi (t) is a fuzzy basis function, and θi (t) is an adjustable parameter of the ˆ ¯ i (t) − $ ¯ i (t) } as small as possible. Let the parameter FLS, which makes sup{ $ estimation error be

θ˜i (t) = θi (t) − θi∗ ,

(32)

ˆ εi (t) = $ ¯ i (t) − $ ¯ i (t).

CE

PT

and let the fuzzy approximation error be (33)

As in the literature [7, 33, 47], it is reasonable to assume that sup{|εi (x(t))|} ≤

AC

ε∗ for some ε∗ > 0. Consequently, we have

ˆ ˆ ˆ $ ¯ i (t) − $i (t) = $ ¯ i (θi (t), x(t)) − $ ¯ i (t) − εi (t) = θ˜iT (t)ψi (t) − εi (t).

Thus the controller can constructed as following:    − uri (t)sign(xi (t)) − µ∗− , xi (t) > 0,   ui (t) = 0, xi (t) = 0,     − u (t)sign(x (t)) + µ∗ , x (t) < 0, ri i i + 8

(34)

(35)

ACCEPTED MANUSCRIPT

with uri = b + θiT (t)ψi (x(t)),

(36)

where b > ∗ .

CR IP T

Remark 3. Owing to the special form of FONNs (see the term −Cx(t) in (24)), the proposed controller (35) is different from that by conventional AFC methods for integer-order nonlinear systems (see [7]). Moreover, one dose not need to know the sign of the system states.

AN US

Remark 4. Although the condition b ∈ (ε∗ , +∞) is needed in the controller design, it is actually difficult to know the exact value of ε∗ . On the other hand, we may see in section 4 that good control performance can be obtained even if b is small (b = 1), which implies that the proposed control method has good robustness. The FOAL is given as C α 0 Dt θi (t)

= γi |xi (t)|ψi (x(t)), θi (0) ≥ 0,

(37)

M

where γi > 0 is the design parameter.

Remark 5. FOAL (37) can also be expressed as:

ED

1 θi (t) = θi (0) + Γ(α)

Z

0

t

$i (t − ξ)α−1 |xi (ξ)|ψi (x(ξ))dξ.

(38)

PT

Remark 6. As the term γi |xi (t)|ψi (x(t)) in (37) is not negative, θ(t) ≥ 0 by Lemma 3. As a result, we have uri ≥ 0.

CE

Theorem 1. Consider the FONN (25) under Assumptions 1-3. The AFC (35) together with the FOAL (37) ensures that lim xi (t) = 0. t→∞

AC

Proof. Multiplying xT (t) to both sides of (29) we have α T T T xT (t)ΛC 0 Dt x(t) = x (t)$(t) − x (t)ΛCx(t) + x (t)ϕ(u(t))

= −xT (t)ΛCx(t) + ≤ −xT (t)ΛCx(t) +

n X

k=1 n X

k=1

9

xk (t)$k (t) + xT (t)ϕ(u(t)) ¯ k (t) + xT (t)ϕ(u(t)). |xk (t)|d$

(39)

ACCEPTED MANUSCRIPT

By (34), we obtain n X

k=1

|xk (t)|d$ ¯ k (x(t)) + xT (t)ϕ(u(t)) − xT (t)ΛCx(t)

= xT (t)ϕ(u(t)) − xT (t)ΛCx(t) + d −d

n X

k=1

|xk (t)|θ˜kT (t)ψk (t) + d

n X

k=1

n X

k=1

|xk (t)|εk (t)

≤ −xT (t)ΛCx(t) + xT (t)ϕ(u(t)) + d

k=1

|xk (t)| − d

n X

n X

k=1

|xk (t)|θkT (t)ψk (t)

|xk (t)|θ˜kT (t)ψk (t).

AN US

+ dε∗

n X

|xk (t)|θkT (t)ψk (x(t))

CR IP T

α xT (t)ΛC 0 Dt x(t) ≤

k=1

(40)

From (35) we know ui ≤ −µ∗− for xi (t) > 0 and ui > µ∗+ for xi (t) ≤ 0. By

(27), we have

(ui (t) + µ∗− )ϕi (ui (t)) ≥ d∗i− u2ri (t) ≥ du2ri (t)

M

for xi (t) > 0, and

ED

(ui (t) − µ∗+ )ϕi (ui (t)) ≥ d∗i+ u2ri (t) ≥ du2ri (t)

(41)

(42)

for xi (t) ≤ 0. Consequently,

PT

− uri (t)x2i (t)sign(xi (t))ϕi (ui (t)) ≥ du2ri (t)x2i (t) = du2ri (t)|xi (t)|2 .

(43)

xi (t)ϕi (ui (t)) ≤ −duri (t)|xi (t)|.

AC

CE

That means, for all xi (t),

10

(44)

ACCEPTED MANUSCRIPT

Substituting (36) and (44) into (40) gives α T xT (t)ΛC 0 Dt x(t) ≤ −duri (t)|xi (t)| − x (t)ΛCx(t) − d

k=1

|xk (t)| + d

= −xT (t)ΛCx(t) − d + d(ε∗ − b)

n X

k=1

n X

k=1

n X

k=1

|xk (t)|θkT (t)ψi (t)

|xk (t)|θ˜kT (t)ψi (t)

|xk (t)|.

(45)

AN US

If we set the design parameter b > ε∗ , then

α T xT (t)ΛC 0 Dt x(t) ≤ −x (t)ΛCx(t) − d

Define

k=1

|xk (t)|θ˜kT (t)ψk (t)

CR IP T

+ dε∗

n X

n X

n X

k=1

|xk (t)|θ˜kT (t)ψk (t).

(46)

n

ν(t) =

1 T 1 X 1 ˜T ˜ x (t)Λx(t) + θ (t)θk (t). 2d 2 γk k

(47)

k=1



n X 1 ˜T C α ˜ 1 T α x (t)ΛC D x(t) + θ (t)0 Dt θi (t). 0 t d γ i i=1 i

ED

C α 0 Dt ν(t)

M

According to Lemma 2, it holds that

(48)

From (32) we have

α =C 0 Dt θi (t).

(49)

PT

C α˜ 0 Dt θi (t)

Substituting (37), (46) and (49) into (48), after some straightforward ma-

CE

nipulators, we have

α 0 Dt V

1 (t) ≤ − xT (t)ΛCx(t). d

(50)

AC

From (50) and Lemma 5, we know that both xi (t) and θ˜i (t) are bounded. Consequently, θi (t) is bounded. Because uri is bounded by (36), ui (t) is bounded.

Note that ΛC > 0, xi (t) converges to 0 asymptotically (i.e., lim xi (t) = 0). t→∞

Remark 7. (1) AFC method is also used to control FOSs in [23], where the α-th q T derivative of a squared Lyapunov 21 eT (t)Λe(t) is given by 21 (C 0 Dt e(t)) Λe(t) +

11

ACCEPTED MANUSCRIPT

1 C q 2 e(t)Λ(0 Dt e(t))

(see Equality (33)). However, this is incorrect, the correct

form is

α T T C α = (C 0 Dt e(t)) e(t) + e (t)0 Dt e(t) ∞ X Γ(1 + α) α−k C k +2 D e(t)C e(t). 0 Dt Γ(1 + k)Γ(1 − k + α) 0 t

CR IP T

C α T 0 Dt 2e (t)e(t)

k=1

(2) Actually, we need not to handle the above complicated infinite series. As shown in Theorem 1, we give a strict proof of the stability of the FONNs based on Lemmas 1 and 2 and the proposed Lemmas 3, 4 and 5.

AN US

Remark 8. In [39], adaptive controller has been proposed for FOSs with input nonlinearities. The comparisons between the results of this paper and that of [39] is included in Table 1.

4. Simulation results

In this section three simulation examples will be given to demonstrate the

M

effectiveness of the proposed method. 4.1. Example 1

ED

In system (24), let n = 3, α = 0.97, fi (xi (t)) = tanh(xi (t)), ci = 1, and

AC

CE

PT

Ii = 0 (i = 1, 2, 3).    2.00 −1.20 0 0.5 0 0       A= 1.71 1.15   2.00 , G =  0 0.5 0 . −4.75 0 1.10 0 0 0.7 The input nonlinearities ϕi (ui (t)) (i = 1, 3) are given as:   (ui (t) − 2)(1 − 0.3 sin(ui (t))), ui (t) > 2.0,    ϕi (ui (t)) = 0, − 2.0 ≤ ui (t) ≤ 2.0,     (u (t) + 2)(0.8 − 0.3 cos(u (t))), u (t) < −2.0. i

i

i

The input nonlinearity ϕ2 (u2 (t)) is chosen as   ui (t) > 5,  (1 − 0.3 sin(u2 (t)))(ui (t) − 5),   ϕ2 (u2 (t)) = 0, − 5 ≤ ui (t) ≤ 5,     (0.8 − 0.3 cos(u (t)))(u (t) + 5), u (t) < −5. i i i 12

(51)

(52)

CR IP T

ACCEPTED MANUSCRIPT

Table 1: Our results and the results in [39].

The results in [39]

Our results

System model

MIMO FOSs, and prior knowledge

MIMO FONNs, and the structure

should be known

can be fully unknown

Control method

Switching adaptive controller

AFC

Gain reduction toler-

Their

be

Their exact values are not needed

ances

known in advance

Design parameters

There are 3n control parameters to

There are n + 1 control parameters

be chosen

to be chosen

Parameters to be up-

There are 3n parameters that

There are nN parameters that

dated

should be updated online

should be updated

The advantages of the control

The advantages of our method: (1)

methods in [39]: (1) fewer param-

better robustness; (2) the exact

eters to be updated, less computa-

values of the gain reduction toler-

tion burden; (2) the controller is

ances can be unknown (it is dif-

easier to design and may be more

ficult to obtain their exact values

convenient for practical use

[7])

should

ED

M

values

AC

CE

PT

Conclusions

exact

AN US

Comparison

13

ACCEPTED MANUSCRIPT

Under these parameters and without the control input (i.e., ui (t) ≡ 0), the

dynamical behavior of system (24) is shown in Figure 2.

3

4

2 x2(t)

x3(t)

2 0

x2(t)

0

−3

−2

−1 x1(t)

(c)

6

−1

1

0

−2 −4

−3

−2

(d)

6

0

1

4

x3(t)

4

−1 x1(t)

AN US

2

x3(t)

1 0

4

2 0

2 0

−3

−2

−1 x1(t)

0

−2 −2

1

M

−2 −4

(b)

4

CR IP T

(a)

0

2

4

x2(t)

ED

Figure 2: Dynamical behavior of system (24) in (a) x1 (t) − x2 (t) − x3 (t), (b) x1 (t) − x2 (t) plane, (c) x1 (t) − x3 (t) plane and (d) x1 (t) − x4 (t) plane.

It is worth mentioning, within all simulation studies, the proposed con-

PT

trol method does not need the knowledge of the FONNs’s model. Indeed, the FONNs’s model is only given for simulation purposes. The FLSs use x1 (t), x2 (t)

CE

and x3 (t) as their inputs. From Figure 2 we know that the system variables distribute on interval [-3, 3]. Thus, for each input variable, we define 3 Gaussian membership functions distribute uniformly on interval [-3, 3]. As a result, there are 3 × 3 × 3 = 27 fuzzy rules are used in the simulation. The Gaussian mem-

AC

bership functions are given in Figure 3. The design parameters are chosen as b = 1, γ1 = γ2 = γ3 = 10, x(0) = [2.0, −2.0, 3.0]T , θi (0) = 0 ∈ R27 .

14

ACCEPTED MANUSCRIPT

1

0.6

0.4

CR IP T

Membership functions

0.8

0.2

0 −3

−2

−1

0 xi(t),i=1,2,3

1

2

3

Figure 3: Membership functions.

AN US

In the simulation, sign(·) is replaced by arctan(10·). Our results are given in Figures 4, 5, and 6. These results indicate that good control performance has been achieved. Figure 4 gives the state variables. Figure 5 presents the control inputs. Noting that the input nonlinearities are given as (51) and (52), the control input u2 (t) tends to 5, and ui (t), i = 1, 3 tend to -2 eventually. Figure 6 depicts the input nonlinearities. It should be mentioned that, although b is set very small, the states converge to 0 rapidly. That is to say that the employed

M

FLSs have good approximation abilities.

x1(t)

ED

3

x2(t) x3(t)

1

PT

State variables

2

0

AC

CE

−1

−2

0

1

2

3

4

5 6 Time(second)

7

Figure 4: The state variables.

15

8

9

10

ACCEPTED MANUSCRIPT

10 u1(t) u2(t) u3(t)

0

−5

−10

0

1

2

3

4

5 6 Time(second)

7

8

6

Φ2(u2(t)) Φ3(u3(t))

2 0 −2

M

Input nonlinearities

10

Φ1(u1(t))

4

−4 −6

0

1

2

3

ED

−8

9

AN US

Figure 5: Control inputs.

CR IP T

Control inputs

5

4

5 6 Time(second)

7

8

9

10

PT

Figure 6: The input nonlinearities.

In the remainder of this subsection, we will show how different choices of membership functions and fuzzy rules affect the control performance. We give

CE

four examples presented in Table 2 (the step is 0.02) by using different membership functions and fuzzy rules. The simulation results are included in Figure 7. In this paper, the fuzzy membership functions are chosen according to expertise

AC

and control performance. It should be mentioned that similar functions are used in [6, 7, 21, 22, 23, 26, 33, 45, 46]. According to simulation results in Figure 7, the proposed AFC works well even when different membership functions are used. It can be concluded that, in all these cases, the simulation results are similar. However, when more fuzzy membership functions are chosen, more fuzzy rules and simulation time are needed. Moreover, the triangular membership function

16

ACCEPTED MANUSCRIPT

computation will consume more time than the Gaussian function. Thus, in the beginning of this subsection, we use three Gaussian functions in the simulation. Further investigation is needed to determine the optimal number of membership

Table 2: Different membership functions

Member functions

Parameters

Fuzzy rules

Time consumed

3 Gaussian functions

[1.2, i], i = 1, 2, 3

27

1.52 seconds

5 Gaussian functions

[0.4, i], i = 1, · · · , 5

125

5 Triangular functions

[i − 2, i, i + 2], i = 1, · · · , 5

125

[0.2, i], i = 1, · · · , 7

343

Case 1: 3 Gaussian functions

4

0

ED 5 Time(second)

2

x1(t) x3(t)

0

0

x3(t)

0 −1 0

5

10

Case 4: 7 Gaussian functions x1(t)

3

x2(t)

−1 −2

x2(t)

1

4

1

CE

State variables

PT

3

x1(t)

Time(second)

Case 3: 5 triangular functions

4

2

−2

10

State variables

−2

State variables

x3(t)

M

State variables

2

−1

51.63 seconds

3

x2(t)

0

10.63 seconds

Case 2: 5 Gaussian functions

4

x1(t)

3

1

7.88 seconds

AN US

7 Gaussian functions

AC

CR IP T

functions and their parameters to achieve better control performance.

x2(t)

2

x3(t)

1 0 −1

5 Time(second)

−2

10

0

5

10

Time(second)

Figure 7: Simulation results for different choices of membership functions.

17

ACCEPTED MANUSCRIPT

4.2. Example 2 Suppose that the FONNs have time-varying gain aij (t) (i.e., the matrix A is replaced by A(t)) and external disturbances [5], which is more complicated

CR IP T

than above model: α = 0.98, fi (xi (t)) = tanh(|x i (t)| − 1), ci = 1, Ii = 0,   a11 (x1 (t)) 1 −9  1.99, |xi (t)| ≤ 1,   , aii (xi (t)) = A= −9 a22 (x2 (t)) 1    −1.99, |x (t)| > 1. i 1 −9 a33 (x3 (t)) To show the robustness of the control system, let us add three external distur-

AN US

bances to system (24): d1 (t) = 0.48 sin t, d2 (t) = −0.48 cos t, d3 (t) = 0.49   sin t − 1.2 −0.2 0.1    0.5 cos t. The control gain matrix is chosen as G =  −0.2 1.2 0.1  , which 0.1 0.1 1.7 is a positive definite matrix. The FLSs has control parameters and input nonlinearities chosen as above. Let x(0) = [−3, 2, −3] and θi (0) = 0 ∈ R27 , i = 1, 2, 3. Our results are presented in Figures 8 and 9. From Figure 8 we can see that

the state variables converge to the origin very fast, and have small fluctuations. Figure 9 depicts the time response of the input nonlinearities. The qualitative

M

analysis of these results is the same as that of Example 1.

1 0

x2(t) x3(t)

PT

State variables

x1(t)

ED

2

−1

CE

−2 −3

AC

−4

0

1

2

3

4

5 6 Time(second)

Figure 8: The state variables.

18

7

8

9

10

ACCEPTED MANUSCRIPT

5 Φ1(u1(t))

0

−5

0

1

2

3

4

5 6 Time(second)

7

CR IP T

Φ3(u3(t))

8

9

10

AN US

Input nonlinearities

Φ2(u2(t))

Figure 9: The input nonlinearities.

4.3. Example 3

It should be mentioned that our controller is also valid for many classes of

M

FOSs. We will give a comparison between our control method and that of [39]. The mathematical model of the uncertain FOS used in [39] (see Equality (76)

ED

on page 218 in [39]) can be described as  C α   0 Dt x1 (t) =10(x2 (t) − x1 (t)) + x4 (t) + 0.2 cos(3x1 (t))x1 (t) + ϕj (u1 (t)),     α  C 0 Dt x2 (t) =28x1 (t) − x2 (t) + x1 (t)x3 (t) + 0.12 sin(4x2 (t))x2 (t) + ϕ2 (u2 (t)), (53)

CE

PT

8 C α    0 Dt x3 (t) =x1 (t)x2 (t) − x3 (t) + 0.12 sin(4x2 (t))x2 (t) + ϕ3 (u3 (t)),  3     C Dα x (t) = − x (t)x (t) − x (t) + 0.1 cos(2x (t))x (t) + ϕ (u (t)), 2 3 4 4 4 4 4 0 t 4

with α = 0.93.

Let the initial condition be x(0) = [2, 4, 1, 3]T . The input nonlinearities

AC

ϕi (ui (t)) (i = 1, 3) are given as (51). The control parameters in [39] are chosen as ci = 0.25, ηi = 8.24, κi λi = 6 (see Equalities (78) and (79) in [39]). Our control parameters are chosen the same as that in Example 2. The simulation results are presented in Figures 10 and 11. The simulation results indicate that under the chosen parameters, compared with the control method in [39], our method has smaller control inputs and faster convergence of the state vector.

19

ACCEPTED MANUSCRIPT

In (53), d(t) = [0.2x1 cos 3x1 , 0.12x2 sin 4x2 , 0.12x2 sin 4x2 , 0.1x4 cos 2x4 ]T . To show the robustness of the proposed control method, let us use larger disturbance: cd(t) where c is a positive constant. The simulation results are shown in Figure 12. It can be concluded that by using the control method in [39],

CR IP T

when the disturbance is chosen as 3d(t), the system variables needs more time

to converge to the origin. When the disturbance is chosen as 6d(t), the controller proposed by [39] will be not valid anymore (to overcome this problem, one can choose larger control parameters). However, our control method works

well when the disturbance changes largely. That is to say, the proposed AFC

(a)

x 2 (t) x 3 (t) x 4 (t)

2 0

0.5

1

1.5

Time(second) (c)

6

x 1 (t)

4 2

0

0.5

1

1.5

x 3 (t) x 4 (t)

2 1

0

Time(second)

x 2 (t) x 3 (t) x 4 (t)

0.4

0.6

x 1 (t)

4

x 2 (t) x 3 (t)

2

x 4 (t)

0 -2

2

0.2

x 1 (t)

Time(second) (d)

6

PT

0

ED

x 2 (t)

(b)

×10 284

3

0

2

State variables

4

State variables

x 1 (t)

6

0

State variables

4

M

State variables

8

AN US

method has satisfactory robustness.

0

0.5

1

1.5

2

Time(second)

CE

Figure 12: Our results and the results in [39] with larger disturbance: (a) the results of [39] when c = 3; (b) the results of [39] when c = 6; (c) our results when c = 3; (d) our results

AC

when c = 6.

5. Conclusions An AFC for a class of FONNs with input nonlinearities and dead-zones has

been proposed in the paper, where the unknown structure of the FONNs and the exact values of gain reduction tolerances are approximated by FLSs. By using the proposed FOAL, the stability of the closed-loop system is proven strictly 20

ACCEPTED MANUSCRIPT

based on fractional Lyapunov criteria. By using the results of this study, many control methods that are valid for integer-order NNs can be extended to control FONNs. Simulation studies have demonstrated that the proposed method is not only efficient for obtaining parameter convergence but also useful for improving

CR IP T

the control performance and robustness. Future work will be focused on how

to choose appropriate fuzzy rules and membership functions to achieve better control performance. 6. Acknowledgements

This work is supported by the National Natural Science Foundation of China

AN US

under Grant 11771263, and the Natural Science Foundation of Anhui Province

of China under Grant 1808085MF181. The authors would like to thank the editor and the anonymous referees for their comments and suggestions. References

M

References

[1] N. Aguila-Camacho, M. A. Duarte-Mermoud, J. A. Gallegos. Lyapunov

ED

functions for fractional order systems. Communications in Nonlinear Science and Numerical Simulation 19 (2014) 2951–2957. [2] W. Ai, W. Chen, J. Xie. A zero-gradient-sum algorithm for distributed co-

PT

operative learning using a feedforward neural network with random weights. Information Sciences 373 (2016) 404–418.

CE

[3] G. A. Anastassiou. Approximations by multivariate perturbed neural network operators. Analysis and Applications 15(2017) 413–432.

AC

[4] M. R. Askari, M. Shahrokhi, M. K. Talkhoncheh. Observer-based adaptive fuzzy controller for nonlinear systems with unknown control directions and input saturation. Fuzzy Sets and Systems, 314 (2017) 24–45.

[5] H. Bao, J. Cao. Projective synchronization of fractional-order memristorbased neural networks. Neural Networks 63 (2015) 1–9.

21

ACCEPTED MANUSCRIPT

[6] A. Boulkroune, A. Bouzeriba, T. Bouden. Fuzzy generalized projective synchronization of incommensurate fractional-order chaotic systems. Neurocomputing 173 (2016) 606-614.

CR IP T

[7] A. Boulkroune. A fuzzy adaptive control approach for nonlinear systems with unknown control gain sign. Neurocomputing 179(2016) 318–325.

[8] W. J. Chang, F. L. Hsu. Sliding mode fuzzy control for takagi–sugeno

fuzzy systems with bilinear consequent part subject to multiple constraints, Information Sciences 327 (2016) 258–271.

[9] H. Chaoui, M. Khayamy, A. A. Aljarboua. Adaptive interval type-2 fuzzy

AN US

logic control for PMSM drives with a modified reference frame. IEEE Transactions on Industrial Electronics, 64 (2017) 3786–3797.

[10] L. Chen, C. Liu, R. Wu, Y. He, Y. Chai, Finite-time stability criteria for a class of fractional-order neural networks with delay. Neural Computing and Applications 27 (2016) 549-556.

[11] M. Hamdy, S. Abd-Elhaleem, M. A. Fkirin. Time-varying delay compensa-

M

tion for a class of nonlinear control systems over network via H∞ adaptive fuzzy controller. IEEE Transactions on Systems, Man, and Cybernetics:

ED

Systems 47 (2017) 2114–2124.

[12] Y. C. Hung, J. J. Yan, T. L. Liao. Projective synchronization of chua’s chaotic systems with dead-zone in the control input, Mathematics and

PT

Computers in Simulation 77 (2008) 374–382. [13] S. Kamal, A. Raman, B. Bandyopadhyay. Finite-time stabilization of fractional order uncertain chain of integrator: an integral sliding mode ap-

CE

proach, IEEE Transactions on Automatic Control 58 (2013) 1597–1602.

[14] E. Kaslik, I.R. Rdulescu. Dynamics of complex-valued fractional-order

AC

neural networks. Neural Networks 89 (2017) 39-49.

[15] A. A. A. Kilbas, H. M. Srivastava, J. J. Trujillo. Theory and applications of fractional differential equations, Elsevier, 2006.

[16] M. P. Lazarevi´c, P. Tzekis. Robust second-order PDα type iterative learning control for a class of uncertain fractional order singular systems. Journal of Vibration and Control 22 (2016) 2004-2018. 22

ACCEPTED MANUSCRIPT

[17] H. Li, Z. Chen, L. Wu, H. K. Lam, H. Du. Event-triggered fault detection of nonlinear networked systems. IEEE Transactions on Cybernetics 47 (2017) 1041–1052.

CR IP T

[18] H. Li, L. Wang, H. Du, A. Boulkroune. Adaptive fuzzy backstepping tracking control for strict-feedback systems with input delay. IEEE Transactions on Fuzzy Systems 25 (2017) 642–652.

[19] Y. Li, Y. Q. Chen, I. Podlubny. Mittag–leffler stability of fractional order nonlinear dynamic systems, Automatica 45 (2009) 1965–1969.

[20] Y. Li, S. Sui, S. Tong. Adaptive fuzzy control design for stochastic nonlinear

AN US

switched systems with arbitrary switchings and unmodeled dynamics. IEEE transactions on cybernetics 47 (2017) 403-414.

[21] Y. Li, S. Tong, L. Liu, G. Feng. Adaptive output-feedback control design with prescribed performance for switched nonlinear systems. Automatica 80 (2017) 225–231.

M

[22] Y. Li, S. Tong, T. Li. Hybrid fuzzy adaptive output feedback control design for uncertain mimo nonlinear systems with time-varying delays and input

ED

saturation, IEEE Transactions on Fuzzy Systems 24 (2016) 841–853. [23] T. C. Lin, C. H. Kuo. H∞ synchronization of uncertain fractional order chaotic systems: Adaptive fuzzy approach, ISA transactions 50 (2011)

PT

548–556.

[24] J. Liu, C. Wu, Z. Wang, L. Wu. Reliable filter design for sensor networks in the type-2 fuzzy framework. IEEE Transactions on Industrial Informatics,

CE

13 (2017) 1742–1752.

[25] J. Liu, S. Vazquez, L. Wu, A. Marquez, H. Gao, L. G. Franquelo. Extended

AC

state observer-based sliding-mode control for three-phase power converters. IEEE Transactions on Industrial Electronics, 64(2017) 22–31.

[26] H. Liu, S. Li, J. Cao, G. Li, A. Alsaedi, F. E. Alsaadi. Adaptive fuzzy prescribed performance controller design for a class of uncertain fractionalorder nonlinear systems with external disturbances, Neurocomputing 219 (2017) 422–430.

23

ACCEPTED MANUSCRIPT

[27] H. Liu, S. Li, G. Li, H. Wang. Adaptive controller design for a class of uncertain fractional-order nonlinear systems: An adaptive fuzzy approach. International Journal of Fuzzy Systems, 20 (2018) 366-379.

CR IP T

[28] H. Liu, S. Li, H. Wang, Y. Huo, J. Luo. Adaptive synchronization for a class of uncertain fractional-order neural networks, Entropy 17 (2015) 7185–7200.

[29] H. Liu, Y. Pan, S. Li, Y. Chen. Synchronization for fractional-order neural networks with full/under-actuation using fractional-order sliding mode

control. International Journal of Machine Learning and Cybernetics (2017)

AN US

doi: 10.1007/s13042-017-0646-z.

[30] H. Liu, Y. Pan, S. Li, Y. Chen. Adaptive fuzzy backstepping control of fractional-order nonlinear systems, IEEE Transactions on Systems Man & Cybernetics: Systems 47 (2017) 2209–2217.

[31] Y. J. Liu, Y. Gao, S. Tong, Y. Li. Fuzzy approximation-based adaptive backstepping optimal control for a class of nonlinear discrete-time systems

M

with dead-zone. IEEE Transactions on Fuzzy Systems 24 (2016) 16–28. [32] Y. J. Liu, Y. Gao, S. Tong, C. P. Chen. A unified approach to adaptive

ED

neural control for nonlinear discrete-time systems with nonlinear dead-zone input. IEEE Transactions on Neural Networks and Learning Systems 27 (2016) 139–150.

PT

[33] Y. Pan, M. J. Er. Enhanced adaptive fuzzy control with optimal approximation error convergence, IEEE Transactions on Fuzzy Systems 21 (2013)

CE

1123–1132.

[34] Y. Pan, Y. Liu, B. Xu, H. Yu. Hybrid feedback feedforward: An efficient design of adaptive neural network control, Neural Networks 76 (2016) 122–

AC

134.

[35] Y. Pan, H. Yu. Biomimetic hybrid feedback feedforward neural-network learning control, IEEE Transactions on Neural Networks and Learning Systems 28 (2017) 1481–1487.

[36] Y. Pan, H. Yu, T. Sun. Global asymptotic stabilization using adaptive fuzzy PD control, IEEE Transactions on Cybernetics 45 (2015) 574–582. 24

ACCEPTED MANUSCRIPT

[37] W. Pedrycz. Fuzzy control and fuzzy systems, Wiley, 1989. [38] I. Podlubny. Fractional differential equations, Academic Press, 1999. [39] M. Roohi, M. P. Aghababa, A. R. Haghighi. Switching adaptive controllers

CR IP T

to control fractional-order complex systems with unknown structure and input nonlinearities. Complexity 21 (2015) 211–223.

[40] J. Shen, J. Lam. Non-existence of finite-time stable equilibria in fractionalorder nonlinear systems, Automatica 50 (2014) 547–551.

[41] A. Theodorakopoulos, G. A. Rovithakis. Guaranteeing preselected tracking

AN US

quality for uncertain strict-feedback systems with deadzone input nonlin-

earity and disturbances via low-complexity control, Automatica 54 (2015) 135–145.

[42] S. Tong, L. Zhang, Y. Li. Observed-based adaptive fuzzy decentralized tracking control for switched uncertain nonlinear large-scale systems with dead zones, IEEE Transactions on Systems, Man, and Cybernetics: Sys-

M

tems 46 (2016) 37–47.

[43] J. C. Trigeassou, N. Maamri, J. Sabatier, A. Oustaloup. A lyapunov ap-

ED

proach to the stability of fractional differential equations, Signal Processing 91 (2011) 437–445.

[44] S. Tyagi, S. Abbas. Stability and synchronization of delayed fractional-

PT

order projection neural network with piecewise constant argument of mixed type. Tbilisi Mathematical Journal 10 (2017) 57–74.

CE

[45] G. Vinodhini, R. M. Chandrasekaran. A comparative performance evaluation of neural network based approach for sentiment classification of online reviews, Journal of King Saud University-Computer and Information Sci-

AC

ences 28 (2016) 2–12.

[46] F. Wang, Z. Liu, Y. Zhang, C. L P. Chen. Adaptive quantized fuzzy control of stochastic nonlinear systems with actuator dead-zone, Information Sciences 370 (2016) 385–401.

[47] L. X. Wang. Adaptive fuzzy systems and control: design and stability analysis, Prentice-Hall, Inc., 1994. 25

ACCEPTED MANUSCRIPT

[48] K. Wiktorowicz. Design of state feedback adaptive fuzzy controllers for second-order systems using a frequency stability criterion. IEEE Transactions on Fuzzy Systems 25 (2017) 499–510.

CR IP T

[49] L. Wu, Y. Gao, J. Liu, H. Li. Event-triggered sliding mode control of stochastic systems via output feedback. Automatica, 82 (2017) 79–92.

[50] T. S. Wu, M. Karkoub, H. S. Chen, W. S. Yu, M. G. Her. Robust tracking

observer-based adaptive fuzzy control design for uncertain nonlinear mimo systems with time delayed states, Information Sciences 290 (2015) 86–105.

[51] Q. Yang, M. Chen. Adaptive neural prescribed performance tracking con-

AN US

trol for near space vehicles with input nonlinearity, Neurocomputing 174 (2016) 780–789.

[52] T. Zhang, X. Xia. Decentralized adaptive fuzzy output feedback control of stochastic nonlinear large-scale systems with dynamic uncertainties, Information Sciences 315 (2015) 17–38.

M

[53] M. Zheng, L. Li, H. Peng, J. Xiao, Y. Yang, H. Zhao. Finite-time projective synchronization of memristor-based delay fractional-order neural networks.

AC

CE

PT

ED

Nonlinear Dynamics 89 (2017) 2641–2655.

26

AC

CE

PT

ED

M

AN US

CR IP T

ACCEPTED MANUSCRIPT

27

Figure 1: Input nonlinearity.

(a)

3

(b)

6

x 2 (t)

AN US

4

x 1 (t)

2

0

2

0

0.5

1

Time(second) (c)

0

2

0

ED

0.5

0.5

1

1.5

0

0.5

1

1.5

2

Time(second) (d)

3

M

1

0

1.5

2

x 4 (t)

1

x 3 (t)

CR IP T

ACCEPTED MANUSCRIPT

1

0

2

0

0.5

1

1.5

2

Time(second)

PT

Time(second)

CE

Figure 10: The time response of the state variables under our control approach (solid line)

AC

and the method of [39] (dashed line) in (a) x1 (t); (b) x2 (t); (c) x3 (t); (d) x4 (t).

28

(a)

50

CR IP T

ACCEPTED MANUSCRIPT

(b)

100

AN US u 2 (t)

u 1 (t)

50 0

0

-50

0

0.5

1

Time(second) (c)

0 -20 -40

0

ED

u 3 (t)

-100

2

0.5

1

1.5

0

0

1

2

Time(second)

PT

Time(second)

2

0

-50

2

1

Time(second) (d)

50

M

20

-60

1.5

u 4 (t)

-50

CE

Figure 11: The time response of the control inputs under our control approach (solid line)

AC

and the method of [39] (dashed line) in (a) u1 (t); (b) u2 (t); (c) u3 (t); (d) u4 (t).

29